The risks of using AI at work have been known for a while, but new evidence keeps piling up. With AI agents that employees set up to perform actions on company computers, the situation only continues to worsen: attackers are already targeting them. Instead of risking detection by using specific malware, attackers can now just send an agent plain text prompts and make it do what they need, such as find secrets on the computer, or send information somewhere external. Every useful skill of an AI assistant becomes an opportunity for the attacker.
It’s not enough for security teams to keep track of which AI tools are in use. They also need to watch the tools’ permissions, instruction sources, connected components, and actual actions. But it doesn’t require adding more security products to do it. Familiar tools can handle many of these tasks: application control, web filtering, EDR, file analysis, and up-to-date threat intelligence. A detailed description of the relevant protection mechanisms can be found in Securelist’s guide to safe deployment of AI agents. That article is packed with technical details, but here, we’ll walk through the main risks of corporate AI, and the traditional security measures that help minimize them.
An assistant nobody approved
Shadow AI refers to AI services and apps that employees use without approval from the company’s security and IT departments. The app or service itself is usually legitimate, but the organization doesn’t know what data it processes, where it sends that data, or what actions it can perform.
To prevent the use of such tools, a registry of approved AI apps and online services should be created and maintained. To restrict access to AI services that security policy doesn’t allow, Kaspersky endpoint security, managed through Kaspersky Security Center can be used. The restriction is applied by the Application Control module, which has an AI Tools category. It can be run in a mode that blocks prohibited apps (denylist), or in a stricter mode that allows only explicitly approved ones (allowlist). Access to web-based AI services is managed by a similar tool: Web Control. For approved services, it should be explicitly specified which types of data can be processed in them. The service providers’ contractual commitments on data protection should also be explicitly checked.
Fake installers instead of AI
This threat is only indirectly related to AI, but we’re mentioning it because attackers loved it so much in 2026. A user searches for a link or installation instructions for an AI agent, clicks an ad, and lands on a page that looks almost identical to the developer’s real website. The page tells the user to copy a command into Terminal — and the result is predictable. In March, we described a version of this scam that uses fake Claude Code and OpenClaw installation pages: instead of a useful tool, users got an infostealer.
Several levels of protection against these attacks — known as ClickFix — should be used. Installation packages and instructions for approved tools should be used through the corporate catalog. Web filtering and anti-phishing mechanisms should help block known fake pages, while endpoint protection should detect malicious payloads and suspicious launch chains.
Excessive agent permissions
Desktop apps, such as the official ChatGPT app, are often seen as the same chatbot in a separate window. In reality, they’re full-fledged AI agents that can work with files and run programs within the permissions they’ve been given. In practice, an agent often has all the privileges of the current user. If action confirmation and isolation are also turned off, a mistaken or malicious instruction can have very serious consequences.
In September 2026, researchers described the hijacking of an active AI assistant session at a SaaS provider. First, the AI agent recommended infected software from the PyPI repository, and then installed it on its own. The attackers got hold of GitHub OAuth tokens and used them to spread the Shai-Hulud worm further.
Last year’s s1ngularity attack is an even more striking example. The malicious script searched for installed AI tools and launched them in a mode that skips confirmation prompts. The agents were told to look for SSH keys, .env files, and crypto wallets. There was no need to hack the model to do this.
This makes a local AI agent an extremely powerful weapon in living-off-the-land attacks. To spot an agent being used for the wrong purpose, its launch parameters, access to credential stores, and child processes should be monitored closely.
To counter this threat, Kaspersky NEXT EDR Optimum users have access to special rules that detect dangerous launch modes. A separate group of rules tracks agent access to files with credentials, tokens, and keys. Depending on the situation, the response can include terminating the AI agent process or even isolating the host.
Dangerous MCP servers
The MCP protocol allows to connect extra tools and data sources to AI applications. For example, an MCP server can search for documents, call APIs, or work with files. However, the sandbox and restrictive policies that apply to the AI agent don’t extend to MCP. That makes connecting an MCP server about as risky as installing a new software component.
In February 2026, cybersecurity experts discovered the SANDWORM_MODE campaign. The McpInject module placed a malicious MCP server in a hidden directory, and added it to the configurations of Claude Code, Cursor, Continue, and Windsurf. The tool descriptions contained injected instructions to search for secrets, such as SSH keys and cloud credentials.
To keep MCP risks under control, an organization needs to maintain an allowlist of MCP servers, monitor changes to their configurations, and limit network connections. Local servers should be vetted like any executable software, and remote ones — like external services with access to corporate information.
Kaspersky EDR has special rules to detect suspicious MCP server scripts that are created or modified in hidden and temporary directories. Such a location doesn’t prove a component is malicious, but it’s a good reason to check where it came from and which process installed it.
Malware that calls LLMs directly
Attackers have learned to use LLMs to generate code and scripts on the fly, requesting what they need straight from the infected system. Google’s reports describe two malware families — PROMPTSTEAL and HONESTCUE — that were found in real-world attacks. The former queried a model through the Hugging Face API to get Windows commands; the latter received C# code from Gemini and compiled it in memory.
From a network monitoring perspective, in both cases the connection goes to a legitimate AI service. That’s why it’s not enough to look at the destination address alone: it’s essential to know which process opens the connection and what it does with the response.
A practical measure is to allow access to model APIs only for the applications and workflows that truly need it, and to investigate any unexpected requests together with endpoint events.
In Kaspersky EDR, two rules can help to spot unusual applications connecting to known external language model APIs and detect local model files. Finding them doesn’t mean a device is infected, but it can help uncover an AI tool that nobody accounted for.
Malicious skills
An agent’s capabilities are often increased with ready-made skills: sets of instructions, scripts, and helper files. Installing them effectively adds one more supply chain that the security team has to manage.
In February 2026, researchers found 341 malicious skills in the ClawHub catalog. Many of them suggested users install so-called necessary components, which, if they did, in actual fact saw them run the AMOS infostealer. Later, Unit 42 described additional ways to bypass security checks that the attackers used, such as inflating the package size to exceed scanner limits.
Skills, along with their related scripts, components, and archives, should be checked before they’re handed over to an agent. This can be done with Kaspersky Scan Engine, built directly into the extension onboarding process, and by adding the standard checks of the package source, version, and contents.
How to limit the impact of an attack
Detection and response are essential, but Secure by Design AI implementation should reduce potential damage even before protection kicks in.
Least privilege. Whenever possible, run the agent under a separate account or in an isolated environment. Limit access to project directories and only the services the agent needs. A ban written as a text instruction is no substitute for isolation.
Credentials outside the agent’s reach. Don’t keep long-lived tokens and keys in configuration files the agent can read. Whenever possible, use a secrets vault, short-lived credentials, and an intermediary service that strictly performs the requested operation without revealing the secret to the agent.
Verified dependencies. Download skills, packages, and models through an internal repository or a controlled proxy. Pin approved versions, and compare checksums against a trusted reference.
Restricted outbound connections. Allow access only to the model APIs, MCP servers, and other resources that the agent really needs. Don’t forget about child processes: blocking the agent’s main process won’t help much if the shell it launches still has unrestricted internet access.
A dedicated response playbook. After a compromise, check the configuration of agents and development environments, event handlers, and connected MCP servers. Revoke affected tokens and active sessions, and replace any keys that may have been exposed.
An approved alternative instead of a flat ban. Give employees an approved AI tool, clear rules for handling data, and a fast process for approving new features. Otherwise, even a policy that looks strict on paper will only push people toward shadow AI.
AI